Looking inside Gomory Aussois, January ISMP Pure Cutting Plane Methods for ILP: a computational perspective Matteo Fischetti, DEI, University of Padova Rorschach test for OR disorders: can you see the tree?

Looking inside Gomory Aussois, January ISMP Outline 1.Pure cutting plane methods for ILPs: motivation 2.Kickoff: Gomory’s method for ILPs (1958, fractional cuts) 3.Bad (expected) news: very poor if implemented naively 4.Good news: room for more clever implementations Based on joint work with Egon Balas and Arrigo Zanette

Looking inside Gomory Aussois, January ISMP Motivation Modern branch-and-cut MIP methods are heavily based on Gomory cuts  reduce the number of branching nodes to optimality However, pure cutting plane methods based on Gomory cuts alone are typically not used in practice, due to their poor convergence properties Branching as a symptomatic cure to the well-known drawbacks of Gomory cuts — saturation, bad numerical behavior, etc. From the cutting plane point of view, however, the cure is even worse than the disease — it hides the trouble source!

Looking inside Gomory Aussois, January ISMP The pure cutting plane dimension Goal: try to come up with a viable pure cutting plane method (i.e., one that is not knocked out by numerical difficulties)… … even if on most problems it will not be competitive with the branch-and-bound based methods This talk: Gomory's fractional cuts (FGCs), for several reasons: –simple tableau derivation –reliable LP validity proof (runtime cut-validity certificate) –all integer coefficients  numerically more stable than their mixed-integer counterpart (GMIs)

Looking inside Gomory Aussois, January ISMP Rules of the game: cuts from LP tableau Main requirement: reading (essentially for free) the FGCs directly from the optimal LP tableau Cut separation heavily entangled with LP reoptimization! Closed loop system (tableau-cut-tableau) without any control valve: highly unstable! Intrinsically different from the recent works on the first closure by F. & Lodi (Chvatal-Gomory closure) and Balas & Saxena and Dash, Gunluk & Lodi (GMI/split closure) where separation is an external black-box decoupled from LP reoptimization

Looking inside Gomory Aussois, January ISMP Bad news: Stein15 (LP bound) Toy set covering instance from MIPLIB; LP bound = 5; ILP optimum = 8 The multi-cut vers. generates rounds of cuts before each LP reopt.

Looking inside Gomory Aussois, January ISMP The tip of the iceberg Bound saturation is just the tip of the iceberg Let’s have a look under the sea… … with our brand-new 3D glasses

Looking inside Gomory Aussois, January ISMP Bad news: Stein15 (LP sol.s) Fractionality spectrography: color plot of the LP sol.s (muti-cut vers.) After few iterations, an almost-uniform red plot (very bad…) iter.t=0t=1t=2t x1x x * 1 (t) x2x x * 2 (t) … … xjxj x * j (t) … …

Looking inside Gomory Aussois, January ISMP Bad news: Stein15 (LP sol.s) Plot of the LP-sol. trajectories for single-cut (red) and multi-cut (blue) versions ( multidimensional scaling) Both versions collapse after a while  no more fuel?

Looking inside Gomory Aussois, January ISMP Bad news: Stein15 (determinants) Too much fuel !!

Looking inside Gomory Aussois, January ISMP Cuts and Pivots Very long sequence of cuts that eventually lead to an optimal integer solution  cut side effects that are typically underestimated when just a few cuts are used within an enumeration scheme A must! Pivot strategies to keep the optimal tableau clean so as generate clean cuts in the next iterations In particular: avoid cutting LP optimal vertices with a weird fractionality (possibly due to numerical inaccuracy)  the corresponding LP basis has a large determinant (needed to describe the weird fractionality)  the tableau contains weird entries that lead to weaker and weaker Gomory cuts

Looking inside Gomory Aussois, January ISMP Role of degeneracy Dual degeneracy is an intrinsic property of cutting plane methods It can play an important role and actually can favor the practical convergence of a cutting plane method… … provided that it is exploited to choose the cleanest LP solution (and tableau) among the equivalent optimal one Unfortunately, by design, efficient LP codes work against us! They are so smart in reducing the n. of dual pivots, and of course they stop immediately when primal feasibility is restored!  The new LP solution tends to be close to the previous one  Small changes in the LP solution imply large determinants  Large determinants imply unstable tableaux and shallow cuts  Shallow cuts induce smaller and smaller LP solution changes  Hopeless!

Looking inside Gomory Aussois, January ISMP Dura lex, sed lex … In his proof of convergence, Gomory used the lexicographic (dual) simplex to cope with degeneracy  lex-minimize (x 0 = c T x, x 1, x 2, …, x n ) Implementation: use a modern LP solver as a black box: –Step 0. Minimize x 0 --> optimal value x * 0 –Step 1. Fix x 0 = x * 0, and minimize x 1 --> optimal value x * 1 –Step 2. Fix also x 1 = x * 1, and minimize x 2 --> optimal value x * 2 –... Key point: at each step, instead of adding equation x j = x * j explicitly… … just fix out of the basis all the nonbasic var.s with nonzero reduced cost  Sequence of fast (and clean) reoptimizations on smaller and smaller degeneracy subspaces, leading to the required lex-optimal tableau Lex-min useful for the convergence proof, but … also in practice?

Looking inside Gomory Aussois, January ISMP Good news #1: Stein15 (LP bound) LP bound = 5; ILP optimum = 8 TB = “Text-Book” multi-cut vers. (as before) LEX = single-cut with lex-optimization

Looking inside Gomory Aussois, January ISMP Good news #1: Stein15 (LP sol.s) TB = multi-cut vers. (as before) LEX = single-cut with lex-optimization Fractionality spectrography

Looking inside Gomory Aussois, January ISMP Good news #1: Stein15 (LP sol.s) Plot of the LP-sol. trajectories for TB (red) and LEX (black) versions (X,Y) = 2D representation of the x-space ( multidimensional scaling)

Looking inside Gomory Aussois, January ISMP Good news #1: Stein15 (determinants) TB = multi-cut vers. (as before) LEX = single-cut with lex-opt.

Looking inside Gomory Aussois, January ISMP Good news #1: sentoy (max. problem) TB = multi-cut vers. (as before) LEX = single-cut with lex-opt.

Looking inside Gomory Aussois, January ISMP Good news #1: sentoy TB = multi-cut vers. (as before) LEX = single-cut with lex-opt. Avg. geometric distance of x* from the Gomory cut

Looking inside Gomory Aussois, January ISMP Good news #1: sentoy Avg. geometric distance between two consecutive optimal sol.s x* TB = multi-cut vers. (as before) LEX = single-cut with lex-opt.

Looking inside Gomory Aussois, January ISMP Ok, it works … but WHY? Enumerative interpretation of the Gomory method (Nourie & Venta, 1982)

Looking inside Gomory Aussois, January ISMP The underlying enumeration tree Any fractional solution x * can be visualized on a lex-tree The structure of the tree is fixed (for a given lex-order of the var.s) Leaves correspond to integer sol.s of increasing lex-value (left to right)

Looking inside Gomory Aussois, January ISMP The “bad” Gomory (TB = no lex) lex-value z may decrease  risk of loop in case of naïve cut purging!

Looking inside Gomory Aussois, January ISMP Nice “sign pattern” of lex-optimal tableau x*0x* X*1X* X*5X* X*kX*k X*8X* X*hX*h X * X 25 XhXh X0X0 RHS  basic var.s   nonbasic var.s  Green row: nonbasic “+” var. x j increases  a basic var x k with k < h increases increasing lex. order  X 10 XjXj

Looking inside Gomory Aussois, January ISMP The key FGC property for convergence Take the tableau row associated with the (lex) first fractional var. x * h where and We want to lex-increase the optimal value  add a FGC in its ≥ form: (a FGC in its ≤ form will not work!). Two cases for the new LP-opt. x [BRANCH] x j = 0 for all j ε J +  [BACKTRACK] otherwise, a “previous component” increases  BIG lex- increase

Looking inside Gomory Aussois, January ISMP The “good” Gomory (lex & ≥)

Looking inside Gomory Aussois, January ISMP A “still bad” Gomory (lex but ≤) … slow sequence, but still monotonically lex-increasing (not enough for finite convergence)

Looking inside Gomory Aussois, January ISMP Lessons learned The Gomory method is framed within its enumerative cast “Good” FGCs may allow for large backtracking steps, but they cannot modify the underlying tree Inefficient depth-first branching on an unnatural variable order  branching even on integer- valued variables!!

Looking inside Gomory Aussois, January ISMP Good news #2: lex on the fly Facts: If x * h is the first fractional var. of the current lex-optimal LP sol., there is no harm in changing the lex sequence from position h Our lex-reoptimization method allows one to do this “natively”, in an effective way The first fractional var. x* h plays the role of the branching var. in enumerative method One can borrow from enumerative methods any clever selection policy for the branching variable x* b (b for branching), and move this var. in the h- th position of the current lex-order  (hopefully) no more branchings on integer variables!

Looking inside Gomory Aussois, January ISMP Variants: get rid of the obj. function The first branching variable x 0 is the objective function  a very unnatural choice for an enumerative method! In some cases, this choice forces Gomory’s method to visit a same subtree several times (see e.g. the Cook-Kannan-Schrijver example below)  Try to get rid of the obj. function: use of invalid cuts (L-CP), binary search, etc. BUT: are these still pure cutting plane methods ?? Let z := 1000 y z integer

Looking inside Gomory Aussois, January ISMP Role of cuts & dynamic lex-order L-CP and L-B&B work on the same underlying tree (L-CP exploiting FGCs) *.dyn versions modify the lex-order on the fly (no branching on integer var.s)

Looking inside Gomory Aussois, January ISMP Computational tests

Looking inside Gomory Aussois, January ISMP Thank you Lex. dual simplex Gomory cuts

Looking inside Gomory Aussois, January ISMP Question: what about GMI cuts? Bits required to represent the integer cut coeff.s when approximating GMI cuts (approx. error =1 for FGCs, approx. error = 0 for GMIs)  GMI cuts appear numerically much more difficult to handle (at least, in a pure cutting plane context …)